Source-linked AI summary
DeepPicker: a Deep Learning Approach for Fully Automated Particle Picking in Cryo-EM
Feng Wang, Huichao Gong, Gaochao liu, Meijing Li, Chuangye Yan, Tian Xia, Xueming Li, Jianyang Zeng
TL;DR
Particle picking is labor-intensive, subjective, and a major obstacle to automated cryo-EM analysis. DeepPicker uses cross-molecule training to automate particle selection, showing robust performance across defocus levels and providing particles useful for downstream cryo-EM structure determination.
Problem
Particle picking requires substantial human effort and can introduce subjectivity and inconsistency, creating a major obstacle for automated cryo-EM pipelines.
Method
DeepPicker trains a deep learning classifier on known particles from previously solved molecular complexes, then refines it using target-molecule particles identified in an initial iteration.
Results
Tests on γ-secretase, spliceosome, and TRPV1 micrographs found relatively robust automated particle-picking performance across defocus levels from 1 µm to 3.5 µm.
Takeaways & Limitations
Automatically picked particles represented a majority of good-quality 2D images and provided a good starting point for 3D map reconstruction.
Abstract
from arXiv · showhide
Particle picking is a time-consuming step in single-particle analysis and often requires significant interventions from users, which has become a bottleneck for future automated electron cryo-microscopy (cryo-EM). Here we report a deep learning framework, called DeepPicker, to address this problem and fill the current gaps toward a fully automated cryo-EM pipeline. DeepPicker employs a novel cross-molecule training strategy to capture common features of particles from previously-analyzed micrographs, and thus does not require any human intervention during particle picking. Tests on the recently-published cryo-EM data of three complexes have demonstrated that our deep learning based scheme can successfully accomplish the human-level particle picking process and identify a sufficient number of particles that are comparable to those manually by human experts. These results indicate that DeepPicker can provide a practically useful tool to significantly reduce the time and manual effort spent in single-particle analysis and thus greatly facilitate high-resolution cryo-EM structure determination.
1 Introduction
High-resolution cryo-EM requires selecting hundreds of thousands of high-quality particles, but manual picking is laborious, subjective, and difficult to automate. DeepPicker addresses this gap with cross-molecule deep learning designed for fully automated particle selection.
- Hundreds of thousands of high-quality particles may be required for high-resolution cryo-EM structure determination.
- Manual particle picking is labor-intensive and time-consuming, requiring substantial human effort to obtain enough good-quality particles.
- Manual selection can introduce subjectivity, bias, and inconsistency as human judgment changes over time.
- Existing computational particle-picking methods include generative, unsupervised, and discriminative approaches, with generative methods often measuring similarity to references.
- DeepPicker was proposed to determine whether deep learning can enable fully automated particle selection without manual intervention.
- DeepPicker uses known particles from previously determined, different molecular complexes to learn cross-molecule features for target-complex picking.
- Tests on TRPV1, human γ-secretase, and yeast spliceosome data found sufficient automated particles comparable to those picked manually by experts.
2 Results
DeepPicker trains a CNN with labeled samples and applies it to micrographs, using cross-molecule data and target-particle refinement for fully automated picking. Across published datasets, the method produced particles comparable to manual selections and supported downstream analysis.
- Automated framework: DeepPicker combines CNN model training with particle picking, using the trained classifier to select particle images from input micrographs.
- Automated framework: The CNN assigns each window a particle probability, producing a scored map of particle likelihoods across the micrograph.
- Fully automated strategy: Cross-molecule training combines known particles from other complexes, then refines the model with target particles identified during the first iteration.
- Fully automated results: Fully automated tests on γ-secretase, spliceosome, and TRPV1 used particles from the other two molecules as training data and compared results with manual references.
- Fully automated results: Recall exceeded 0.81 across fully automated tests, and recall and precision were relatively high when selected-particle counts approached the reference counts.
- Fully automated results: Automated particles were comparable to manual selections despite differences in micrograph contrast and particle shape and size, indicating near-human-level selection.
- Downstream validation: 2D clustering and class averaging produced results comparable to manually selected particles, providing a good starting point for 3D map reconstruction.
- Center accuracy: Automatically picked particle centers deviated less than 10% of the sliding-window box size from manually identified centers.
3 Discussion
DeepPicker’s fully automated cross-molecule strategy performs robustly across defocus conditions and avoids straightforward particle-like selections from pure-noise micrographs. Its offline training and GPU-enabled implementation support an efficient, near human-level particle-picking tool.
- Robustness: Across defocus levels from 1 µm to 3.5 µm, the fully automated method showed relatively robust performance on γ-secretase, spliceosome, and TRPV1 micrographs.The study examined recall scores at different defocus levels.
- Noise testing: On 26 pure-background micrographs, the model found it difficult to pick particles using the normal threshold used for experimental datasets.The background images were measured from an empty cryo-EM instrument without a sample.
- Training strategy: Cross-molecule training makes picking objective, can incorporate particles from previously solved structures, and may increase training robustness.The framework uses computational expertise learned from training data rather than subjective manual picking.
- Training strategy: Offline cross-molecule training can finish before current-target micrograph collection, making it more efficient than online training in many existing classification-based frameworks.The framework can train before collecting data for the target molecule.
- Computational advantages: GPU computing speeds learning, and the authors report that fully automated picking can be generally more efficient and robust than previous semi-automated methods.This comparison is presented as a general advantage of the framework.
- Overall performance: Tests on γ-secretase, spliceosome, and TRPV1 showed near human-level particle-picking performance for both fully automated and semi-automated schemes.The methods identified sufficient particles comparable to those manually selected by human experts.
4 Materials and methods
The study uses CNN-based particle classification, trained with positive and negative samples, and evaluates automated picking against expert-verified manual references. It also specifies cross-molecule training, iterative refinement, preprocessing, scoring, and evaluation procedures.
- Data and evaluation: The datasets were split into non-overlapping training and test sets, with 100 test micrographs each for γ-secretase, spliceosome, and TRPV1.Training data also included micrographs from β-galactosidase and NSF complex datasets.
- CNN model: The CNN classifies sliding-window patch images as positive particles or negative random-noise images using a two-unit softmax output.The architecture contains four ReLU convolutional layers with max pooling, two fully connected layers, and dropout in the first fully connected layer.
- Training strategy: Training uses balanced positive and negative samples, with positives drawn either from known particles of other molecules or manually selected target particles.Negative samples come from regions at least 0.6 sliding-window sizes from positive samples.
- Training strategy: Fully automated training starts with 10,000 particles from four other molecules and iteratively adds the top 10,000 predicted particles as new training data.The approach also tests training combinations containing one to four molecule types.
- Data and evaluation: Performance is measured against expert-verified manual references using recall and precision, with particle agreement defined by a center-distance threshold of 0.2 sliding-window sizes.The study also measures normalized center deviations and compares 2D clustering and class averaging with manual picks.
- Picking pipeline: The pipeline applies Gaussian filtering, CNN scoring, threshold-based cleaning, spatial filtering, bad-particle removal, and iterative model refinement.The sliding-window step size is 4 pixels; the window is 180 pixels for TRPV1 and γ-secretase and 320 pixels for spliceosome.
- Implementation: With a NVIDIA Quadro K4000 GPU, training on 10,000 samples took less than an hour and processing a 3710 × 3838-pixel micrograph took about 1.5 minutes.DeepPicker can run with or without GPUs.
6 Additional information
The paper reports institutional and grant support and identifies the contributors responsible for project initiation, algorithm design, implementation, experiments, data acquisition, and manuscript preparation.
- Funding: The work received support from Chinese national research programs, the National Natural Science Foundation of China, and structural biology initiatives.The listed support includes grants to multiple authors and institutions.
- Author contributions: Authors divided responsibilities across project initiation, program and algorithm design, implementation, testing, data acquisition, and manuscript drafting and revision.JZ and XL initiated the project, while FW, HG, GL, and ML performed testing and analysis.
Par!cles Training data: Molecule B
Figures 5 and 6 examine semi-automated and fully automated particle-picking behavior across datasets, training choices, defocus levels, and random-noise backgrounds. Table 1 lists alternative combinations of training data used for Figure 2 tests.
- Figure 5: Figure 5 compares semi-automated training with fully automated training across γ-secretase, spliceosome, and TRPV1 datasets.Semi-automated CNNs use 400, 1000, or 10000 manually picked target particles; fully automated training uses 10000 particles from four other molecules.
- Figure 5: Figure 5b summarizes recall above prediction score 0.5, while Figures 5c–e plot precision and recall against the number of picked particles.Solid and dashed lines represent recall and precision, respectively.
- Table 1: Table 1 records different combinations of training data corresponding to the fully automated picking tests shown in Figure 2.